Changes in RC4

[fletchjp]

I wonder if someone would summarise for us the changes in RC4 so that we know where to direct testing.

I have already noticed that a directory called quark has disappeared and there is a new one called control.

I notice that the makefile still included references to quark as a source of headers to include, which is presumably redundant.

Thanks

John

[mateo70]

Hi,

I won't enter into details, but quark should have not appear into the previous release. The new directory control contains all the workaround of MAGMA and will contains the future functions for initialisation of MAGMA library in a context for using different cards for example or multi-threaded code like it appeared in RC3.
So because this code was too much "work in progress" I removed it from the release and won't be part of the final 1.0.0.

The others changes are the final integration of the eigenvalues solvers and a lot of cleanings in the f2c code used in these solvers and other small changes like :
- consistency of types: cuDoubleComplex and cuFloatComplex are used everywhere now.
- consistency of the interface of all the functions.
- try to clean most of the return values in lapack functions
- finish to clean interface and code of eigenvalues problem
- integration of some bug report for windows port: multiple definition of min, max, problem with int64_t, problem with threads under windows but not used anymore for this release.

Mathieu

[fletchjp]

Thanks.

John

[fletchjp]

Something has changed in the FORTRAN interface. Calls to magma_dgetrf_gpu and magma_dgetrs_gpu don't compile for rc4. The type of an argument has changed from a pointer to real.

Yes. Looking in testing_dgetrf_gpu_f.f at the definition of the variable used to allocate GPU memory:

In RC3

Code: Select all

      integer(kind=16)              :: devptrA, devptrB


In RC4

Code: Select all

      real, dimension(4)            :: devptrA, devptrB


This has broken my code. Making the above change restores it.

Also, the change means that something which did not work now works. Hooray!!

I was getting nan instead of results, and now it works.

John

[mateo70]

Yes , sorry I forgot this change, since I didn't do it.
I have to check why it doesn't work with the size of a pointer (integer kind=8 on 64bits system or integer kind=4 for 32bits system) which should be the right choice for this. And we change it to this value because some compiler were not happy with the kind=16.

Mathieu

[fletchjp]

It makes it impossible to do pointer arithmetic on it - see my other post on this.

John